Source-linked AI summary
Delineate Anything v2: A Global Foundation Model for Field Delineation
Mykola Lavreniuk, Nataliia Kussul, Andrii Shelestov, Yevhenii Salii, Volodymyr Kuzin, Charlotte Julia Li-Xing Wang, Zoltan Szantoi
TL;DR
Accurate large-scale agricultural field boundary mapping is important for land-management and monitoring tasks, but geospatial models face parcel-merging, weak-boundary, and scale-awareness challenges. Delineate Anything v2 addresses these issues with data-centric supervision and globally representative resources, achieving 0.284 mAP@0.5 (+103.3% relative gain) over Delineate Anything while mapping Ukraine nationwide in 5.4 hours.
Problem
Large-scale field delineation lacks reliable handling of merged administrative parcels and visually weak physical boundaries, despite its importance for agricultural monitoring and land-management decisions.
Method
Delineate Anything v2 uses resolution-aware data curation, FBIS-73M with 73 million instances across 61 countries, and an independent 100-country benchmark.
Results
0.284 mAP@0.5 (+103.3% relative gain) over Delineate Anything, with nationwide Ukraine mapping completed in 5.4 hours on a consumer-grade workstation.
Takeaways & Limitations
The findings suggest that large-scale geospatial foundation-model progress may depend more on improving supervision quality, consistency, and representativeness than on increasing model complexity.
Takeaways & Limitations
FBIS-73M has a heterogeneous, long-tailed sample distribution shaped by open cadastral-data availability rather than deliberate sampling.
Abstract
from arXiv · showhide
Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities, they frequently fail in geospatial domains due to topological complexity, cropland texturing patterns, and a lack of physical scale awareness. In this work, we introduce Delineate Anything v2, a globally scalable foundation model designed specifically for wide-area field boundary mapping. We construct FBIS-73M, a 73-million-instance multi-resolution dataset spanning 61 countries. To address the pervasive issue of multi-field administrative parcel merging, we introduce a resolution-specific data curation pipeline that leverages topological image-space adaptation to homogenize merged parcels and strengthen weak physical boundaries. Furthermore, we establish a novel, manually curated evaluation benchmark covering 100 countries to assess independent zero-shot generalization. Our results show that Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework, by 0.284 mAP@0.5 (+103.3% relative gain), while maintaining execution speeds suitable for rapid national- and global-scale deployment, as demonstrated by nationwide mapping of Ukraine (603,000 km^2) in 5.4 hours on a consumer-grade workstation. Code, pre-trained weights, the FBIS-73M dataset, and ready-to-use national-scale vector boundary products are publicly available at https://lavreniuk.github.io/Delineate-Anything/.
1 Introduction
Delineate Anything v2 addresses global agricultural field delineation by remediating parcel-merging label noise through resolution-specific data curation and image-space adaptation. It introduces a 61-country dataset and 100-country benchmark, surpasses Delineate Anything by 0.284 mAP@0.5 (+103.3% relative gain), and maps Ukraine nationwide in 5.4 hours.
- Motivation: Field boundaries link remote sensing to land-management decisions, crop-type identification, agricultural statistics, food-security assessment, deforestation compliance, and climate-resilience evaluation.The paper presents precise agricultural field contours as a primary spatial unit for these applications.
- Problem: Supervised models perform strongly locally but generalize poorly beyond regions with digital cadastral registries, while general foundation models struggle with geospatial imagery.The introduction contrasts local ground-truth dependence with the limitations of general-purpose vision foundation models such as SAM.
- Method: Delineate Anything v2 remediates parcel-merging label noise using manual HR polygon splitting and automated MR image-space pixel homogenization.The resolution-specific strategy erases false internal boundaries and synthetically enhances valid but visually weak physical edges in MR samples.
- Results and scalability: 0.284 mAP@0.5 (+103.3% relative gain) is the improvement over the Delineate Anything framework, and Ukraine nationwide mapping covered 603,000 km2 in 5.4 hours.The deployment used a consumer-grade workstation and is presented as evidence of rapid national- and global-scale suitability.
2 Related Work
Agricultural field delineation has progressed from traditional image-analysis methods to semantic, instance-based, and foundation-model approaches, but existing methods remain limited by boundary ambiguity, annotation scarcity, geospatial-scale failures, and noisy administrative labels. Dataset coverage and supervision quality therefore remain central challenges for reliable global generalization.
- Traditional methods: Traditional edge detection, watershed segmentation, and object-based image analysis scale poorly for agricultural field-boundary mapping.Field boundaries support yield estimation, crop classification, environmental mapping, subsidy validation, and compliance monitoring.
- Deep learning methods: Semantic segmentation models such as U-Net and ResUNet-a introduced scalable learning, while multi-task formulations targeted boundary proximity, distance maps, and field interiors.Despite boundary-aware losses, semantic networks remain limited by unresolved field-margin ambiguity.
- Instance-based methods: Detection and instance-segmentation models enforce discrete field identities, but their development remains bottlenecked by limited instance-level agricultural annotations.Representative approaches include Co-DETR and real-time YOLO variants, alongside ViT-Adapter and EVP backbone advances.
- Foundation models: Foundation-model adaptations enable zero-shot remote-sensing segmentation, yet lack of geospatial awareness and scale constraints causes over-segmentation or failures on low-contrast boundaries.Existing adaptations use prompt tuning, multi-scale spatial attention, or weakly supervised boundaries.
- Datasets and label quality: FBIS-22M and Delineate Anything expanded training data, but European concentration and parcel-versus-field label noise still constrain reliable global generalization.Earlier datasets also have limited geographic coverage or imagery resolution, including 439K instances in Vietnam/Cambodia, 2.5M parcels in Europe, and 1.6M parcels across 24 countries at 10m imagery.
- Datasets and label quality: Data-centric learning addresses supervision quality at its source, and Delineate Anything v2 targets parcel-versus-field noise using visual heterogeneity and resolution-specific remediation.This contrasts with conventional label-noise methods focused mainly on semantic corruption, wrong class assignments, or annotation errors.
3 Methodology
Delineate Anything v2 uses data-centric curation, combining the 73-million-instance FBIS-73M repository with a manually annotated, 100-country benchmark. Its pipeline detects merged administrative parcels and adapts curation by resolution to suppress false internal boundaries while enhancing weak physical edges.
- Data assets: FBIS-73M contains 73 million agricultural field instances across 61 countries, using high-resolution (0.25–3 m) and medium-resolution (3–10 m) imagery.The training distribution is heterogeneous and long-tailed because it reflects cadastral-data availability, while low-resource regions broaden geographic diversity.
- Data assets: The independent benchmark samples exactly four manually annotated image patches per country across 100 countries to evaluate zero-shot cross-country generalization with equal geographic weighting.It spans agricultural systems from large industrial fields to fragmented smallholder landscapes and emphasizes structural diversity.
- Curation challenges: Administrative parcels can merge multiple physical fields, while weak contrast between adjacent similar-crop fields motivates an automated curation pipeline.These challenges are especially problematic when neighboring fields share similar crops or phenological stages.
- Merged-parcel detection: The pipeline screens instances using blurred, eroded parcel cores and adaptive color-deviation thresholds, flagging candidates when deviation-level areas exceed 15%, 10%, or 5% of parcel area.Baseline threshold sets are [10%, 20%, 50%] for Sentinel-2 and [25%, 35%, 50%] for Planet imagery.
- Resolution-specific adaptation: For HR imagery (≤3 m GSD), flagged parcels receive manual geometric decomposition; for MR imagery (3 −10 m GSD), image-space adaptation homogenizes pixels while preserving texture and reducing internal boundaries.The cyclical bounding transformation uses B = 32 for Sentinel-2 and B = 48 for Planet imagery, while topology-based edge enhancement targets weak true boundaries.
4 Results
Delineate Anything v2 achieves state-of-the-art zero-shot field delineation across a 100-country benchmark, with strong regional gains and improved separation of fragmented fields. Data-centric remediation contributes more than dataset scaling, while inference remains suitable for nationwide mapping.
- Global benchmark: Delineate Anything v2 establishes state-of-the-art performance on the independent 100-country zero-shot benchmark, doubling DelAny’s mAP@0.5 (+0.284) and nearly tripling mAP@0.5:0.95 (+0.175).It also delivers a +0.294 gain in precision.
- Regional performance: Performance doubles in Africa (+0.333) and nearly triples in Asia & Oceania (+0.279), while Europe and North America exceed 0.61 mAP@0.5.The model bridges performance differences between mechanized agriculture and fragmented smallholder systems.
- Ablation study: Scaling the training corpus from 22M to 73M instances yields only a +0.086 baseline expansion before structural label noise causes a performance ceiling.The ablation supports supervision quality as a larger bottleneck than dataset scale.
- Data-centric remediation: High-Resolution manual geometric splitting adds a critical +0.048 mAP by resolving multi-field parcel merging and preventing topological collapse in fragmented smallholder landscapes.Qualitative results likewise show sharp separation in ultra-fragmented landscapes and low-contrast boundaries, unlike existing layers that often collapse or merge fields.
5 Conclusion
Delineate Anything v2 presents a data-centric foundation model for global agricultural field boundary delineation, emphasizing label quality as the central bottleneck. Its large-scale dataset, independent benchmark, strong global performance, and public releases support scalable and reproducible agricultural monitoring.
- Contributions: Delineate Anything v2 is introduced as a foundation model built on a data-centric supervision paradigm for agricultural field boundary delineation.The conclusion identifies label quality as the fundamental bottleneck at global scale.
- Contributions: FBIS-73M contains 73 million field instances across 61 countries and uses resolution-aware remediation to correct parcel-versus-field mismatch.The dataset is described as the largest publicly available field boundary repository to date.
- Evaluation: The independent benchmark covers 100 countries and diverse agricultural systems, enabling evaluation of global generalization outside the training distribution.The benchmark was curated specifically for independent assessment beyond the training distribution.
- Results: 603,000 km^2 of Ukraine was mapped nationwide in 5.4 hours on a consumer-grade workstation, demonstrating scalable deployment.The conclusion describes this result as part of a new state of the art in global field delineation.
- Resources: The authors release FBIS-73M, the 100-country benchmark, model weights, code, and global vector products for reproducible research and large-scale agricultural monitoring.Applications include food security, environmental compliance, and planetary-scale Earth observation.